In Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy
will defend his dissertation
Automatic Characterization of Stories
Computerized systems capable of generating high-level story descriptions have many potential real-life applications. However, enabling computers to do so requires teaching computers to obtain an abstract understanding of natural language stories algorithmically, which is one of the non-trivial problems in Artificial Intelligence and Natural Language Processing. In this thesis, we tackle the challenge of automatically characterizing stories at a high-level by generating a set of tags from narrative texts written in natural language. We focus on narratives written in English and explore two directions for solving this problem. First, we hypothesize that by designing and applying abstraction-based methods on narrative texts, we would be able to achieve a high-level understanding of stories and describe a story with multiple tags. Second, we argue that it is possible to take advantage of user reviews of story-based items to extract story descriptions and construct tagsets from those descriptors. We make several contributions to tackle the challenge of automatic story characterization through this thesis. In the first part, we present a background study on the problem, discuss required resources for research, and propose a new corpus to facilitate research on high-level story understanding by selecting tag predictions for movies as an application of this problem. Then in the second part, we focus on designing methods for high-level story understanding from written narratives and experiment on predicting tags for movies from the written plot synopses. We first employ a wide range of linguistic features to design a machine learning approach for generating descriptive tags for stories from narrative texts. At the next step, we design a neural methodology for modeling the flow of emotions throughout stories and enhance a system that uses a high-level representation of narrative texts to predict tags. Finally, we exploit the hierarchical structure of text documents to encode the synopses for employing in a tag-prediction model and improve the tag-prediction performance. In the final part of this thesis, we explore how we can utilize user reviews to generate tags for characterizing stories at a high level. We begin by proposing a multi-view learning paradigm for retrieving story attributes from narrative texts and user reviews simultaneously. Then we develop a method to extract story descriptors from reviews as tags without any direct supervision. We made the new dataset, source code of systems, and a live tag prediction system publicly available to the community to encourage further exploration in the direction of automatic story characterization.
Date: Monday, April 13, 2020
Time: 10:00 - 11:30 AM
Place: Online Presentation - Zoom Meeting
Advisor: Dr. Thamar Solorio
Faculty, students, and the general public are invited.